The price of thinking.

A token is the unit that AI is sold by, and the price of one has fallen further and faster than almost anything in the history of technology. Except when it has not. Here is what a thought actually costs in September 2026, why the number is so slippery, and the one habit that separates a $14,000 bill from a $200 one: stop sending every task to the smartest model. Two calculators inside price your own workload, and your own routing.

Scroll

Prologue

Act I / the menu

What every frontier model charges, September 2026.

Act ii / the collapse

The price of a fixed amount of intelligence is in free fall.

Cost of gpt-4-class intelligence over time

Cheapest blended $ per million tokens to reach ~gpt-4 capability. Log scale. Hover for the model.

Act iii / intelligence per dollar

The real question is not what a model costs. It is what it costs to be smart.

Intelligence vs price, today's frontier

Capability (ctg index) against blended $ per million tokens, log price axis. Up and to the left wins. Hover any model.

Act iv / price your own thinking

What does your workload actually cost?

Input tokens / task

Output tokens / task

Tasks / month

Act v / route your traffic

Why is your best model answering your easiest questions?

Frontier model · for the hard, undefined work

Workhorse model · for the understood, repeatable work

Share of your tasks that are routine enough to route down

Act vi / why the number is slippery

Seven ways the sticker price lies.

Act vii / the subsidy

The price you pay and the price it costs are not the same number.

Act viii / the European line

And where does Europe price its thinking?

The ctg. take

Epilogue / one number

Notes / methodology